add distort and space

configs/rec/rec_chinese_lite_train.yml

@@ -14,6 +14,7 @@ Global:
   character_type: ch
   character_dict_path: ./ppocr/utils/ppocr_keys_v1.txt
   loss_type: ctc
+  distort: true
   reader_yml: ./configs/rec/rec_chinese_reader.yml
   pretrain_weights:
   checkpoints:

configs/rec/rec_icdar15_train.yml

@@ -13,6 +13,7 @@ Global:
   max_text_length: 25
   character_type: en
   loss_type: ctc
+  distort: true
   reader_yml: ./configs/rec/rec_icdar15_reader.yml
   pretrain_weights: ./pretrain_models/rec_mv3_none_bilstm_ctc/best_accuracy
   checkpoints:

ppocr/data/rec/dataset_traversal.py

@@ -45,6 +45,8 @@ class LMDBReader(object):
         self.use_tps = False
         if "tps" in params:
             self.ues_tps = True
+        if "distort" in params:
+            self.use_distort = params['distort']
         if params['mode'] == 'train':
             self.batch_size = params['train_batch_size_per_card']
             self.drop_last = True
@@ -142,7 +144,8 @@ class LMDBReader(object):
                                 label=label,
                                 char_ops=self.char_ops,
                                 loss_type=self.loss_type,
-                                max_text_length=self.max_text_length)
+                                max_text_length=self.max_text_length,
+                                distort=self.use_distort)
                             if outs is None:
                                 continue
                             yield outs
@@ -185,6 +188,8 @@ class SimpleReader(object):
         self.use_tps = False
         if "tps" in params:
             self.use_tps = True
+        if "distort" in params:
+            self.use_distort = params['distort']
         if params['mode'] == 'train':
             self.batch_size = params['train_batch_size_per_card']
             self.drop_last = True
@@ -232,9 +237,14 @@ class SimpleReader(object):
                         img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
 
                     label = substr[1]
-                    outs = process_image(img, self.image_shape, label,
-                                         self.char_ops, self.loss_type,
-                                         self.max_text_length)
+                    outs = process_image(
+                        img=img,
+                        image_shape=self.image_shape,
+                        label=label,
+                        char_ops=self.char_ops,
+                        loss_type=self.loss_type,
+                        max_text_length=self.max_text_length,
+                        distort=self.use_distort)
                     if outs is None:
                         continue
                     yield outs

ppocr/data/rec/img_tools.py

@@ -15,6 +15,7 @@
 import math
 import cv2
 import numpy as np
+import random
 from ppocr.utils.utility import initial_logger
 logger = initial_logger()
 
@@ -89,6 +90,252 @@ def get_img_data(value):
     return imgori
 
 
+def flag():
+    """
+    flag
+    """
+    return 1 if random.random() > 0.5000001 else -1
+
+
+def cvtColor(img):
+    """
+    cvtColor
+    """
+    hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
+    delta = 0.001 * random.random() * flag()
+    hsv[:, :, 2] = hsv[:, :, 2] * (1 + delta)
+    new_img = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
+    return new_img
+
+
+def blur(img):
+    """
+    blur
+    """
+    h, w, _ = img.shape
+    if h > 10 and w > 10:
+        return cv2.GaussianBlur(img, (5, 5), 1)
+    else:
+        return img
+
+
+def doudong(img):
+    """
+    doudong
+    """
+    w, h, _ = img.shape
+    if h > 10 and w > 10:
+        thres = min(w, h)
+        s = int(random.random() * thres * 0.01)
+        src_img = img.copy()
+        for i in range(s):
+            img[i:, i:, :] = src_img[:w - i, :h - i, :]
+        return img
+    else:
+        return img
+
+
+def add_gasuss_noise(image, mean=0, var=0.1):
+
+    noise = np.random.normal(mean, var**0.5, image.shape)
+    out = image + 0.5 * noise
+    out = np.clip(out, 0, 255)
+    out = np.uint8(out)
+    return out
+
+
+def get_crop(image):
+    """
+    random crop
+    """
+    h, w, _ = image.shape
+    top_min = 1
+    top_max = 8
+    top_crop = int(random.randint(top_min, top_max))
+
+    crop_img = image.copy()
+
+    ratio = random.randint(0, 1)
+    if ratio:
+        crop_img = crop_img[top_crop:h, :, :]
+    else:
+        crop_img = crop_img[0:h - top_crop, :, :]
+    return crop_img
+
+
+class Config:
+    """
+    Config
+    """
+
+    def __init__(self, ):
+        self.anglex = random.random() * 30
+        self.angley = random.random() * 15
+        self.anglez = random.random() * 10
+        self.fov = 42
+        self.r = 0
+        self.shearx = random.random() * 0.3
+        self.sheary = random.random() * 0.05
+        self.borderMode = cv2.BORDER_REPLICATE
+
+    def make(self, w, h, ang):
+        """
+        make
+        """
+        self.anglex = random.random() * 5 * flag()
+        self.angley = random.random() * 5 * flag()
+        self.anglez = -1 * random.random() * int(ang) * flag()
+        self.fov = 42
+        self.r = 0
+        self.shearx = 0
+        self.sheary = 0
+        self.borderMode = cv2.BORDER_REPLICATE
+        self.w = w
+        self.h = h
+
+        self.perspective = True
+        self.crop = True
+        self.affine = False
+        self.reverse = True
+        self.noise = True
+        self.dou = False
+        self.blur = True
+        self.color = True
+
+
+def rad(x):
+    """
+    rad
+    """
+    return x * np.pi / 180
+
+
+def get_warpR(config):
+    """
+    get_warpR
+    """
+    anglex, angley, anglez, fov, w, h, r = \
+        config.anglex, config.angley, config.anglez, config.fov, config.w, config.h, config.r
+    if w > 69 and w < 112:
+        anglex = anglex * 1.5
+
+    z = np.sqrt(w**2 + h**2) / 2 / np.tan(rad(fov / 2))
+    # Homogeneous coordinate transformation matrix
+    rx = np.array([[1, 0, 0, 0],
+                   [0, np.cos(rad(anglex)), -np.sin(rad(anglex)), 0], [
+                       0,
+                       -np.sin(rad(anglex)),
+                       np.cos(rad(anglex)),
+                       0,
+                   ], [0, 0, 0, 1]], np.float32)
+    ry = np.array([[np.cos(rad(angley)), 0, np.sin(rad(angley)), 0],
+                   [0, 1, 0, 0], [
+                       -np.sin(rad(angley)),
+                       0,
+                       np.cos(rad(angley)),
+                       0,
+                   ], [0, 0, 0, 1]], np.float32)
+    rz = np.array([[np.cos(rad(anglez)), np.sin(rad(anglez)), 0, 0],
+                   [-np.sin(rad(anglez)), np.cos(rad(anglez)), 0, 0],
+                   [0, 0, 1, 0], [0, 0, 0, 1]], np.float32)
+    r = rx.dot(ry).dot(rz)
+    # generate 4 points
+    pcenter = np.array([h / 2, w / 2, 0, 0], np.float32)
+    p1 = np.array([0, 0, 0, 0], np.float32) - pcenter
+    p2 = np.array([w, 0, 0, 0], np.float32) - pcenter
+    p3 = np.array([0, h, 0, 0], np.float32) - pcenter
+    p4 = np.array([w, h, 0, 0], np.float32) - pcenter
+    dst1 = r.dot(p1)
+    dst2 = r.dot(p2)
+    dst3 = r.dot(p3)
+    dst4 = r.dot(p4)
+    list_dst = [dst1, dst2, dst3, dst4]
+    org = np.array([[0, 0], [w, 0], [0, h], [w, h]], np.float32)
+    dst = np.zeros((4, 2), np.float32)
+    # Project onto the image plane
+    for i in range(4):
+        dst[i, 0] = list_dst[i][0] * z / (z - list_dst[i][2]) + pcenter[0]
+        dst[i, 1] = list_dst[i][1] * z / (z - list_dst[i][2]) + pcenter[1]
+    warpR = cv2.getPerspectiveTransform(org, dst)
+
+    dst1, dst2, dst3, dst4 = dst
+    r1 = int(min(dst1[1], dst2[1]))
+    r2 = int(max(dst3[1], dst4[1]))
+    c1 = int(min(dst1[0], dst3[0]))
+    c2 = int(max(dst2[0], dst4[0]))
+
+    try:
+        ratio = min(1.0 * h / (r2 - r1), 1.0 * w / (c2 - c1))
+
+        dx = -c1
+        dy = -r1
+        T1 = np.float32([[1., 0, dx], [0, 1., dy], [0, 0, 1.0 / ratio]])
+        ret = T1.dot(warpR)
+    except:
+        ratio = 1.0
+        T1 = np.float32([[1., 0, 0], [0, 1., 0], [0, 0, 1.]])
+        ret = T1
+    return ret, (-r1, -c1), ratio, dst
+
+
+def get_warpAffine(config):
+    """
+    get_warpAffine
+    """
+    anglez = config.anglez
+    rz = np.array([[np.cos(rad(anglez)), np.sin(rad(anglez)), 0],
+                   [-np.sin(rad(anglez)), np.cos(rad(anglez)), 0]], np.float32)
+    return rz
+
+
+def warp(img, ang):
+    """
+    warp
+    """
+    h, w, _ = img.shape
+    config = Config()
+    config.make(w, h, ang)
+    new_img = img
+
+    if config.perspective:
+        tp = random.randint(1, 100)
+        if tp >= 50:
+            warpR, (r1, c1), ratio, dst = get_warpR(config)
+            new_w = int(np.max(dst[:, 0])) - int(np.min(dst[:, 0]))
+            new_img = cv2.warpPerspective(
+                new_img,
+                warpR, (int(new_w * ratio), h),
+                borderMode=config.borderMode)
+    if config.crop:
+        img_height, img_width = img.shape[0:2]
+        tp = random.randint(1, 100)
+        if tp >= 50 and img_height >= 20 and img_width >= 20:
+            new_img = get_crop(new_img)
+    if config.affine:
+        warpT = get_warpAffine(config)
+        new_img = cv2.warpAffine(
+            new_img, warpT, (w, h), borderMode=config.borderMode)
+    if config.blur:
+        tp = random.randint(1, 100)
+        if tp >= 50:
+            new_img = blur(new_img)
+    if config.color:
+        tp = random.randint(1, 100)
+        if tp >= 50:
+            new_img = cvtColor(new_img)
+    if config.dou:
+        new_img = doudong(new_img)
+    if config.noise:
+        tp = random.randint(1, 100)
+        if tp >= 50:
+            new_img = add_gasuss_noise(new_img)
+    if config.reverse:
+        tp = random.randint(1, 100)
+        if tp >= 50:
+            new_img = 255 - new_img
+    return new_img
+
+
 def process_image(img,
                   image_shape,
                   label=None,
@@ -96,7 +343,10 @@ def process_image(img,
                   loss_type=None,
                   max_text_length=None,
                   tps=None,
-                  infer_mode=False):
+                  infer_mode=False,
+                  distort=False):
+    if distort:
+        img = warp(img, 10)
     if infer_mode and char_ops.character_type == "ch" and not tps:
         norm_img = resize_norm_img_chinese(img, image_shape)
     else:

ppocr/utils/ppocr_keys_v1.txt

@@ -6620,4 +6620,5 @@ j
 緖
 續
 紹
-懮
\ No newline at end of file
+懮
+